Carlo S. Regazzoni

Particle PHD Filtering for Multi-Target Visual Tracking

We propose a multi-target tracking algorithm based on the probability hypothesis density (PHD) filter and data association using graph matching. The PHD filter is used to compensate for miss-detections and to remove noise and clutter. This filter propagates the first order moment of the multi-target posterior (instead of the full posterior) to reduce the growth in complexity with the number of targets from exponential to linear. Next the filtered states are associated using graph matching. Experimental results on face, people and vehicle tracking show that the proposed multi-target tracking algorithm improves the accuracy of the tracker, especially in cluttered scenes.

A bio-inspired system model for interactive surveillance applications

Advances in computer vision and pattern recognition research are leading to video surveillance systems with improved scene analysis capabilities. However, up to now few works have handled the problem of how the system, along with a human operator, can actively cope with detected anomalous events. In this paper, on the basis of recent studies on artificial cognitive systems, a general framework is proposed for designing interactive, adaptable and intelligent surveillance systems. The aim of the system is to react to situations in a preventive way using actuators installed in the monitored environment. An application of the proposed system is introduced where a guard is supported in pursuing an intruder. The operator is first localized and tracked and then multi-modal guidance messages are communicated to him on a mobile device. Previous experience on the interaction dynamics between the two players is provided by a simulator, modeling guard and intruder behaviors, to predict near future events and decide the appropriate messages to be sent. Results on real world video sequences show the reliability of the simulated data to build up interaction models and predict near future events. Moreover, the system capability of learning relationships with the operator to establish efficient and personalized communications is verified.

Neural Networks Mode Classification based on Frequency Distribution Features

The growing number of new emerging wireless standards is creating regulatory problems in allocating the unlicensed frequencies. A possible solution for increasing the frequency re-usage within the framework of info-mobility cellular systems is the joint exploitation of Smart Antennas and Cognitive Radio. In the paper a Mode Identification algorithm, based on frequency distribution features and multiple neural network classifiers, for a Cognitive Base Transceiver Station is presented. Simulated results, obtained in a simplified framework, will prove the effectiveness of the proposed approach.

Multimodal cognitive system for immersive user interaction

In the recent years many efforts have been made to provide machines with the capability of effectively interact with its users. Here a system is introduced that supplies a virtual guide service to users by means of a mobile device (e.g. Palm, tablet PC, etc.). The architecture takes inspiration from a biological model of the cognitive processes, the Cognitive Cycle, performed by the brain while interacting with other entities. A variety of multimodal interfaces communicates messages to interact adaptively with the user. Results show the effectiveness of the different message modalities in real situation where an user is moving towards a target guided by the system.

HOS-based mode classification for infomobility framework

The growing number of new emerging wireless standards is creating regulatory problems in allocating the unlicensed frequencies. A possible solution for increasing the frequency reusage within the framework of info-mobility cellular systems is the joint exploitation of Smart Antennas and Cognitive Radio. Inside this framework a key-role is played by Mode Identification and Spectrum monitoring algorithms, useful to provide awareness about the channel conditions. In the paper a Mode Identification algorithm, based on the extraction of higher order statistics from frequency distribution of the involved communication modalities and multiple support vector machine classifiers, for a Cognitive Base Transceiver Station is presented. Simulated results, obtained in a simplified framework, will prove the effectiveness of the proposed approach.

Bio-inspired Cognitive Radio for Dynamic Spectrum Access

Dynamic spectrum access (DSA) has raisedthe attention of industrial and academic researchers due to the fact thatit is seen as a technologyable to overcome the lack of available spectrum for new communication services.In particular, autonomic DSA (ADSA) systems are indicated as a solution to spectrumscarcity caused by the current “command and control” allocationparadigm. However, ADSA requires a higher level of reconfigurability with respect totraditional wireless systems. In this context, one of the technologies thatcan provide such flexibility is the promising cognitive radio (CR).In an ADSA scenario, CR should sense the spectrum to find the resources unused byprimary (licensed) users, which could then be exploited by secondary(unlicensed) CR users to increase the overall system efficiency.In this chapter, a comprehensive overview of CR applications to ADSA is carried out;in particular, attention is paid to the potentialities of autonomic bio-inspiredapproaches, and on their advantages in the solution of the challenges ofADSA systems.

Comparing the Performance of Learnable Evolution Model LEM and Pattern Search as a Function Optimizer

The underlying paper presents a comparison of the learnable evolution model LEM and Pattern Search PS techniques as a function optimizer. In contrast to conventional Darwinian type evolutionary computation algorithm that uses various forms of mutation and/or recombination operators, LEM uses machine learning to guide the process of generating new individuals. It employs the AQ learning to generate hypotheses discriminating between groups of high and low fitness individuals, and then uses these hypotheses to generate new individuals. On the other hand pattern search is a class of direct search for derivative-free optimization with accurately established global convergence properties. Pattern search makes no use of derivative information, which might be unavailable, too expensive, or misleading. This paper focuses on measuring the performance of LEM3 and pattern search from the point of view of execution time in experiments on optimizing the Rastrigin function with different number of variables.